Large language models featuring weights from 9B to 397B parameters are now available under an MIT license from the Ornith team. The Ornith-1.5 models use an end-to-end self-improvement loop in which the system generates its own tasks and solution rollouts for reinforcement learning. This approach allows the models to reach state of the art performance among open models for coding and agentic tasks.
Inference platforms vLLM, Ollama, and SGLang integrated serving support for the models immediately following the launch. The vLLM and SGLang stacks currently support the 9B variant, while the full series is accessible via Ollama. Quantized versions in FP8, GGUF, MLX, and NVFP4 formats were released to facilitate deployment across diverse hardware.
Key sources
- SOURCE@vllm_project“SOTA among open models on coding and agentic tasks”x.com
- SUPPORT@ornith_“vllm serve ornith-ai/Ornith-1.5-9B”x.com
- SOURCE@ornith_“Ornith 1.5 series models are now available on ollama!”x.com
- SUPPORT@sgl_project“thanks for including SGLang in the model cards”x.com
- SOURCE@ornith_“Now you can serve Ornith 1.5 with SGLang!”x.com
- SOURCE@ornith_“developed on top of Qwen3.5 @Alibaba_Qwen with additional continued pretraining, mid-training, and post-training”x.com
- SUPPORT@ornith_“AD-Q4_K runs on a 16GB MacBook Air with 64k context and picks the same next token as the BF16 original 91.9% of the time”x.com
- SUPPORT@ornith_“This is a 9B model running a full agentic loop with Wi-Fi off”x.com